Google DeepMind's Gemini Robotics 2, released on July 30, 2026, splits robot intelligence into three specialized models so a machine can THINK, ACT, and ADAPT independently. The thinking model (ER 2) plans multi-step tasks and coordinates entire robot fleets; the action model (Robotics 2 VLA) translates those plans into physical motion from feet to fingertips; and the efficiency model (On-Device 2) runs locally and adapts to entirely new robot bodies in just a few hours using fewer than 200 demonstration examples. Together they represent Google's most practical architecture for real-world deployment: the reasoning layer is publicly available to developers today on the Gemini API and Google AI Studio, while the action and on-device layers are in early-access partner programs.
Last verified: 2026-08-01 — ER 2 is live on Gemini API and AI Studio; VLA and On-Device 2 require partner early-access.
- Bold takeaway: A generalist "brain" now orchestrates any low-level "body."
- Fast body swap: New robots in ~a few hours via On-Device 2 motion transfer, <200 examples.
- Safety leap: 97.9% instruction-following, 93.0% human-proximity detection (vs ER 1.6's 47.2% and 51.1%).
- Multi-robot: ER 2 demonstrated a humanoid + arm + wheeled robot collaborating on shared tasks.
- Developer-ready: ER 2 available in AI Studio; VLA/On-Device need partner waitlist. Volatile facts: model availability and access tiers are subject to change as partner programs expand.
What Is Gemini Robotics 2 and Why Split the Brain From the Body?
Gemini Robotics 2 is a three-tier model stack—two VLA models and one VLM—released by Google DeepMind on July 30, 2026. The key design insight is DECOUPLING: you do not need one model that does everything. Instead, a reasoning model breaks a task into steps and calls execution models the way a software agent calls an API. This mirrors what successful multi-agent orchestration systems already do in the software world, a pattern we've covered in our deep dive on Sakana Fugu's multi-agent orchestration.
The important architectural change is that the reasoning model no longer needs to know how a robot's motors work. It just moves the reasoning overhead off motor-control-decision time, and you can plug in a different motor executor underneath. That means a humanoid from Apptronik, a Spot quadruped from Boston Dynamics, and a tabletop arm from Franka Robotics can share the same high-level brain, each running a different low-level action model beneath it—just as a human brain does not care whether the body is tall or short; it just decides "lift the cup."
The three models:
| Model | Role | Google DeepMind describes it as | Runs where |
|---|---|---|---|
| Gemini Robotics 2 (VLA) | Vision-language-action — maps perception + language to motor commands | "Most advanced VLA model" — full humanoid control, feet to fingertips | Robot's compute unit |
| Gemini Robotics ER 2 (VLM) | Embodied reasoning — plans, tracks progress, coordinates multiple robots | "High-level brain for robots" | Cloud via Gemini Live API (sub-second streaming) |
| Gemini Robotics On-Device 2 (VLA, efficient) | Local VLA — offline adaptation to new robot bodies | "Most efficient VLA, optimized to run locally" | onboard robot hardware |
The split is a big deal because previous VLA models tried to reason and act in a single network. That meant an upgrade to reasoning couldn't deploy without retraining the action layer—a constraint that slowed real-world deployment. By formally separating reasoning (ER 2) from action execution (Robotics 2 VLA), Google's architecture lets engineers change one layer without disturbing the other—which is the core promise of any modern agent operating system: architecture should be modular so the system can evolve in pieces.
How Does Gemini Robotics ER 2 Plan Multi-Step Tasks?
Gemini Robotics ER 2 handles long-horizon task planning by streaming continuous video from the robot's cameras, breaking complex instructions into ordered steps, tracking task completion across five progress levels (0-20%, 20-40%, 40-60%, 60-80%, 80-100%), and calling execution tools (VLA models or robot APIs) to carry out each step.
The controller of the system only understands text in and text out. It receives multimodal input (video, audio, text), reasons, and emits text instructions to the motor layer. The critical engineering detail is how ER 2 runs. It integrates with the Gemini Live API, Google's bidirectional streaming endpoint optimized for latency-sensitive tasks. This means the robot can "think" about the next step while simultaneously performing the current one—eliminating the stop-and-think pauses that made earlier models unusable for real-time work.
Here's the actual whole-body demo Google published: an Apptronik Apollo 2 humanoid is told, "put the watering can into the green bin in the bottom shelf." The robot walks to the table, picks up the watering can, takes a few steps to the shelves, and places it precisely in its destination. Apollo processes the instruction, plans the multi-step sequence from walking through grasping through navigating shelving to placing, and completes the task from a single natural-language instruction. There's no human joystick. The humanoid is operating under one connected policy that controls the legs, torso, arms, and fingers—what Google calls "whole-body" control, a step-change from Robotics 1.5's upper-body-only manipulation.
What benchmark numbers did ER 2 actually hit?
The ER 2 numbers are from Google DeepMind's own internal benchmarks, reported in the official release blog post. These are vendor-reported results, not independent third-party evaluation.
| Evaluation | ER 1.6 (previous gen) | ER 2 | Best competitor (OR @line) |
|---|---|---|---|
| Success detection (image) | 82.9% | 87.7% | 83.6% (Opus 5) |
| Success detection (video) | 76.0% | 82.4% | 81.0% (Opus 5) |
| Question answering (ERQA) | 72.5% | 78.5% | 73.0% (Gemini 3.6 Flash) |
| Generalized instrument reading | 52.8% | 65.7% | 61.5% (GPT 5.6 Sol) |
| Progress classification | 42.7% | 57.4% | 46.2% (GPT 5.6 Sol) |
Source: Google DeepMind ER 2 model page and Google Blog announcement.
Two things to note about the progress classification score. First, at 57.4% accuracy, models still get it right less than six times in ten; this is the clearest signal of how far remains on the "when is the step done?" problem that determines whether a robot can truly operate autonomously. Second, the comparison suite includes Opus 5 and GPT 5.6 Sol, heavyweight general-purpose models—ER 2 wins on every measured eval despite being a robotics-specialized model based on Gemini 3.5 Flash, the efficient tier.
What is Moment-Finding precision?
Moment-finding is the task of identifying the exact video frame where something critical happens—like knowing the precise moment to stop pouring coffee so the cup doesn't overflow. ER 2 achieves 91.3% accuracy and a 0.96-second mean absolute distance (it lands within roughly one second of the correct frame, on average). Google reports this performance at a fraction of the compute cost and 4x the execution speed of larger models, meeting the sub-second latency required for safe physical operation.
Can Gemini Robotics On-Device 2 Really Adapt to a New Robot Body in Hours?
Yes, for bi-arm embodiments. Gemini Robotics On-Device 2 uses what DeepMind calls "motion transfer"—a technique inherited from Gemini Robotics 1.5—to adapt to new robot hardware with fewer than 200 demonstration examples collected over a few hours. This is the breakthrough that matters for fast iteration.
DeepMind demonstrated adaptation across three robot platforms not in its training data: the Dexmate, SO101, and Trossen bi-arm systems. Each has a different shape, different sensors, and different degrees of freedom. The fine-tuning runs over a few hours and uses fewer than 200 examples regardless of whether the new embodiment has substantially different joint configurations from what the model was originally trained on.
What it means: A hardware vendor can plug a new robot into the adaptation pipeline, collect roughly 200 demonstrations of a human guiding the arms through movements, and a usable policy comes back without retraining from scratch. The implication is huge: you don't need Google-scale compute to get a working policy for a new robot. The samples needed to adapt are within reach of a small lab.
Honest limitations of On-Device 2
The publicly accessible model card lists key boundaries:
- On-Device 2 is based on Gemini Robotics 1.5 technology plus on-device Gemma models, not the newer Robotics 2 VLA generation.
- It was developed mainly for standing bi-arm manipulation.
- Google says it has limited ability to generalize to out-of-distribution tasks or control high-degree-of-freedom robots.
Multi-finger dexterity success rates for the Robotics 2 VLA range from 32-92% across tasks, with screwing in a lightbulb at only 36%—showing real frontier challenges remain in fine manipulation.
Source: DeepMind On-Device 2 model page and Gemini Robotics 2 DeepMind blog post.
How Does Multi-Robot Collaboration Work in ER 2?
Gemini Robotics ER 2 is the first model in Google's robotics line explicitly designed to coordinate multiple robots on the same task. Different machines—a wheeled rover and a humanoid, for example—communicate via a shared semantic understanding to hand off and complete subtasks. There is no hardcoded handoff logic, because the reasoning model can see the scene and break the work into steps that fit each robot's strengths.
The concrete demo is the clearest signal: DeepMind showed ER 2 directing an Apptronik Apollo 2 humanoid and a Franka F3 Duo arm working together on a shared cleaning task. In a separate demonstration, Google showed the system directing a Boston Dynamics Spot quadruped to retrieve a snack on a natural-language command—wheeling or walking the right machine into the right position to complete a job without human intervention.
Google also demonstrated ER 2 being used to orchestrate Boston Dynamics Spot's existing APIs (navigation, manipulator movement) to build an interactive robot that fetches objects on natural-language commands. Getting started code for the Spot integration is available in the google-gemini/robotics-samples GitHub repository.
For a deep dive on how the underложения multi-agent orchestration pattern parallels what we're seeing in software agents, see our guide on Sakana Fugu's multi-agent orchestration.
Is multi-robot teamwork production-ready?
Not yet. The demo collaborations were conducted in controlled settings. The step-change here is the model can coordinate multiple robots with different capabilities, but real-world fleet deployment requires identity, permissions, shared maps, scheduling, collision avoidance, network resilience, and clear recovery rules when one robot fails. Nobody ships multi-robot collaboration in an uncontrolled environment yet; nobody has solved what happens when the Wi-Fi drops mid-handoff. The honest framing is that Google has made the coordination problem tractable at the model level—the operational plumbing is still the customer's problem.
What Safety Benchmarks Did Google Introduce With Gemini Robotics 2?
Google DeepMind introduced the ASIMOV-Agentic benchmark and released a Safety Technical Report (PDF) accompanying the Gemini Robotics 2 launch. The benchmark evaluates whether a foundation model can act as a safe VLA orchestrator by testing its capacity to enforce safety constraints, monitor the environment, assess physical feasibility, and seek human clarification when uncertain.
The safety scores represent the most dramatic improvement in this release:
| Safety Metric | ER 1.6 | ER 2 | Opus 5 | GPT 5.6 Sol |
|---|---|---|---|---|
| Safety Instruction Following | 47.2% | 97.9% | 95.9% | 91.4% |
| Human Proximity (1m) | 51.1% | 93.0% | 77.1% | 83.4% |
Source: Google DeepMind ER 2 model page.
The jump from ER 1.6 to ER 2 on safety is the largest single-generation improvement in the release. ER 1.6's 47.2% instruction-following score was concerning; ER 2's 97.9% brings it from worst-in-class to best-in-class. The human-proximity score (51.1% to 93.0%) means the model now correctly identifies when a person enters the robot's workspace and reliably halts the robot—a capability Google demonstrated on the Apptronik Apollo 2 humanoid in a garage environment, where a human actor encroached on the workspace from various angles and ER 2 orchestrated a safe stop, settling in-hand objects and moving the robot to a safe pose before autonomously resuming the task after the human departed.
The ASIMOV benchmark family (v1 and v2) was developed by Google DeepMind and Princeton University and is open-source, published at asimov-benchmark.github.io. The new ASIMOV-Agentic extension tests agentic safety orchestration and uncertainty resolution: whether the model refuses unsafe VLA tool calls, predicts infeasibility, and asks humans instead of guessing.
Critical safety caveat from the model card
The ER 2 model card explicitly prohibits use in safety-critical applications: "Users are required to use discretion before using the Robotics Models in a production, commercial, or public environments, and to not use the Robotics Models for safety-critical applications or work." Prohibited use cases listed include healthcare, transportation, and any area where malfunction could foreseeably lead to death, personal injury, or property damage. This is a research preview, not a deployment-ready safety system.
Source: Gemini Robotics ER 2 Model Card.
How Can Developers Access Gemini Robotics ER 2 Today?
Gemini Robotics ER 2 is publicly available as of July 30, 2026 through three channels:
- Google AI Studio — Start a new chat with the
gemini-robotics-er-2-previewmodel. This is the easiest entry point for experimentation. - Gemini API — Full API access documented at ai.google.dev/gemini-api/docs/robotics-overview. Supports text, image, video, and audio input; text output only; up to 128K context window.
- Gemini Enterprise Agent Platform — Private preview access for production-grade orchestration tooling.
The VLA models (Gemini Robotics 2 and On-Device 2) are NOT in general access—they require partner early-access. Developers interested in the motor-control layer need to apply through the waitlist.
Google published getting-started examples including a live demo pairing ER 2 with Boston Dynamics Spot's movement APIs to fetch objects on natural-language commands, with code on GitHub.
The approach mirrors how the broader AI ecosystem is moving toward agentic upgrades: models are no longer just text generators—they're reasoning engines that call tools. ER 2 is one of the first models that explicitly calls robot APIs as tools.
What This Means for You
If you're a robotics integrator or OEM: The ER 2 reasoning model is available today in AI Studio. Try it as your planning layer and declare your existing robot APIs as callable tools. The Live API's bidirectional streaming endpoint is your integration path—stream multimodal video from your robot and receive text instructions back in real time. For new platforms, On-Device 2's <200-example adaptation claim means you can prototype a new embodiment without retraining from scratch—but budget for safety controls, certification, and the limits noted in the model card.
If you're a builder using AI for your work: The three-layer split (reason + act + onboard) is becoming the template for physical AI. The same separation is already how the best software agents work, a shift we traced in our guide on what comes after RLHF. For non-robotics builders, the relevant takeaway is tool orchestration as a first-class capability—ER 2 can call Google Search, calendars, and any user-defined function.
If you're evaluating humanoid robot deployments: Don't confuse demo success with deployment readiness. The benchmark numbers are vendor-reported internal evaluations. The model card explicitly prohibits safety-critical use. Multi-finger dexterity success rates are 32-92%. Progress classification is 57.4%. These are real numbers that mean real robots will still misjudge task completion and fail at fine manipulation in novel scenarios. The architecture is sound; the maturity is not.
FAQ
Q: What is Gemini Robotics ER 2? A: Gemini Robotics ER 2 is Google DeepMind's embodied reasoning model for robots, released July 30, 2026. It acts as a high-level brain that plans multi-step tasks, tracks progress, coordinates multiple robots, and calls lower-level vision-language-action (VLA) models for motor execution. It does not control motors directly.
Q: How fast can Gemini Robotics On-Device 2 adapt to a new robot? A: On-Device 2 adapts to a new bi-arm robot embodiment in a few hours using fewer than 200 demonstration examples, according to Google DeepMind. This was demonstrated on Dexmate, SO101, and Trossen platforms that were not in the original training data.
Q: Is Gemini Robotics ER 2 available to the public? A: Yes. ER 2 is live on the Gemini API and Google AI Studio as of July 30, 2026, with private preview access on the Gemini Enterprise Agent Platform. The VLA and On-Device 2 models require partner early-access.
Q: Can Gemini Robotics 2 control a full humanoid robot? A: Yes. The Gemini Robotics 2 VLA model controls full bipedal humanoids from feet to fingertips, demonstrated on Apptronik's Apollo 2 which walked, picked up objects, navigated shelving, and placed items from a single natural-language instruction. Previous generations controlled only upper-body tabletop manipulation.
Q: How much does Gemini Robotics ER 2 cost? A: Google has not published specific pricing for Gemini Robotics ER 2 as of August 2026. It is available through the Gemini API (subject to Gemini API terms) and Google AI Studio. Check the Gemini API pricing page for current rates.
Q: What is the ASIMOV-Agentic benchmark? A: ASIMOV-Agentic is a safety benchmark introduced with Gemini Robotics 2 that evaluates whether a foundation model can act as a safe VLA orchestrator by testing enforcement of safety constraints, environmental monitoring, physical feasibility assessment, and human clarification seeking. It was developed by Google DeepMind and Princeton University.
Discussion
0 comments